Blog
Research, model releases, and engineering notes from The Token Company.
Introducing bear-2-finance: A Query-Aware, Finance-Tuned Compression Model
Now in private preview: give bear-2-finance a filing and a question and it returns only the relevant spans at a budget you set. On FinanceBench, readers show no measurable loss at 50% of the uncompressed token count.
August 2026
Compressing Conversational Context Without Losing the Thread
Bear-2 compression improved CoQA accuracy from 93.3% to 95.3% while cutting tokens by 8.2%. Why meeting transcripts and multi-turn conversations compress better than you'd expect.
June 2026
Introducing Bear-2-Safety
Token-deletion compression for safety classification pipelines. Cut classifier costs by 30% while preserving or improving F1 across Llama-Guard, ShieldGemma, and Gemini.
June 2026
One of the biggest token consumers globally improved quality by removing context bloat
Processing 193B tokens/month, Pax Historia ran a 268K-vote model arena. Compressed models scored higher and A/B tests showed +5% purchase lift.
February 2026
Helonic: cutting inference costs while maintaining quality
How Helonic uses token compression to reduce LLM API costs without sacrificing output quality.
February 2026